Building Reversible 1d Convolutional Networks in Pytorch

Hello All,

I am trying to build a reversible convolutional network for 1d audio signals. I have 0 clue where to start with this, as I assume just defining an additional boolean in the forward() method won’t work too well because specifically I am trying to train from 2 pairs of losses: 1 derived from going forward in the network with an input signal (loss is backpropogated), and 1 going backward in the network (loss in a sense is forward propogated).

How would y’all recommend implementing a network like this? Should I custom design a reversible 1d convolution class? Or can I get away with adding a parameter to the forward function?

Any help would be much appreciated :slight_smile: